Local Search for Planning and Scheduling

Withtheincreasingdeploymentofplanningandschedulingsystems,developers oftenhavetodealwithverylargesearchspaces,real-timeperformancedemands, anddynamicenvironments. Completere?nementmethodsdonotscalewell,- kinglocalsearchmethodstheonlypracticalalternative. Adynamicenvironment alsopromotestheapplicationoflocalsearch,thesearchheuristicsnotnormally beinga?ectedbymodi?cationsofthesearchspace. Furthermore,localsearchis wellsuitedforanytimerequirementsbecausetheoptimizationgoalisimproved iteratively. Suchadvantagesareo?setbytheincompletenessofmostlocalsearch methods,whichmakesitimpossibletoprovetheinconsistencyoroptimalityof thesolutionsgenerated. Popularlocalsearchapproachesincludeevolutionary- gorithms,simulatedannealing,tabusearch,min-con?icts,GSAT,andWalksat. The?rstarticleinthisbook-aninvitedcontributionbyStefanVoß-givesan overviewofthesemethods. ThebookisbasedonthecontributionstotheWorkshoponLocalSearchfor Planning&Scheduling,heldonAugust21,2000atthe14thEuropeanCon- renceonArti?cialIntelligence(ECAI2000)inBerlin,Germany. Theworkshop broughttogetherresearchersfromtheplanningandschedulingcommunitiesto explorethesetopicswithrespecttolocalsearchprocedures. Aftertheworkshop, asecondreviewprocessresultedinthecontributionstothepresentvolume. Voß’soverviewisfollowedbytwoarticles,byHamiezandHaoandGerevini andSerina,onspeci?c'classical'combinatorialsearchproblems. Thearticleby HamiezandHaoaddressestheproblemofsports-leaguescheduling,presenting results achieved by a tabu search method based on a neighborhood of value swaps. GereviniandSerina’sarticleaddressesthetopicthatdominatestherest ofthebook:actionplanning. Itbuildsontheirpreviousworkonlocalsearch onplanninggraphs,presentinganewsearchguidanceheuristicwithdynamic parametertuning. Thenextsetofarticlesdealwithplanningsystemsthatareabletoinc- porateresourcereasoning. The?rstarticle,ofwhichIamtheauthor,makesit clearwhyconventionalplanningsystemscannotproperlyhandleplanningwith resourcesandgivesanoverviewoftheconstraint-basedExcaliburagent’spl- ningsystem,whichdoesnothavetheserestrictions. Thenextthreearticlesare aboutNASAJPL’sASPEN/CASPERsystem. The?rstone-byChien,Knight, andRabideau-focusesonthereplanningcapabilitiesoflocalsearchmethods, presentingtwoempiricalstudiesinwhichacontinuousplanningprocessclearly outperformsarestartstrategy. Thenextarticle,byEngelhardtandChien,shows howlearningcanbeusedtospeedupthesearchforaplan. Thegoalisto?nda setofsearchheuristicsthatguidethesearchaswellaspossible. Thelastarticle inthisblock-byKnight,Rabideau,andChien-proposesanddemonstrates, a technique for aggregating single search moves so that distant states can be reachedmoreeasily. VI Preface Thelastthreearticlesinthisbookaddresstopicsthatarenotdirectlyrelated tolocalsearch,butthedescribedmethodsmakeverylocaldecisionsduringthe search. RefanidisandVlahavasdescribeextensionstotheGRTplanner,e. g. ,a hill-climbingstrategyforactionselection. Theextensionsresultinmuchbetter performancethanwiththeoriginalGRTplanner. Thesecondarticle-byO- india, Sebastia, and Marzal - presents a planning algorithm that successively re?nes a start graph by di?erent phases, e. g. , a phase to guarantee comp- teness. Inthelastarticle,HiraishiandMizoguchipresentasearchmethodfor constructingaroutemap. Constraintswithrespecttomemoryandtimecanbe incorporatedintothesearchprocess. Iwishtoexpressmygratitudetothemembersoftheprogramcommittee, whoactedasreviewersfortheworkshopandthisvolume. Iwouldalsoliketo thank all those who helped to make this workshop a success - in